Maisy Dunlavy

Advisor: Zhiling Lan

Mentor: Walter Hopkins

Undergraduate: University of New Mexico (Computer Science)

Graduate: University of Illinois Chicago (Computer Science)

Project: Anomaly Analysis Pipeline for Cross-Experimental AIDQM

This project develops a reproducible anomaly analysis pipeline to support data validation and anomaly detection research for the Cross Experimental AIDQM initiative in high energy physics (HEP). The work focuses on improving how anomaly detection models are executed, evaluated, and diagnosed using real and controlled datasets. The pipeline supports anomaly labeling through controlled injection, tracking of true positives (TP), false positives (FP), and false negatives (FN), clustering of false positive events, and automated generation of diagnostic artifacts including CSV outputs and visualization plots. The system has been designed to improve reproducibility and maintainability of anomaly detection experiments. Key features include confusion metric exports, feature-difference ranking, cluster summaries, TP/FP visualizations, train-if-missing model execution, GPU-ready PBS run scripts, and standardized logging and plotting outputs. The pipeline also maintains compatibility with both legacy workflows and a newer anomaly-injection stack, enabling continued use of existing analysis tools while integrating updated components. In addition, the project establishes an experimental infrastructure that enables systematic comparison of models, seeds, and threshold configurations. This allows researchers to evaluate model behavior across controlled experiment grids rather than relying on individual one-off runs. Current work focuses on reducing false positives by identifying persistent FP-heavy clusters and feature signatures and applying targeted threshold tuning, model configuration adjustments, and improved validation reporting.